Given a library of \(N\) \(^1\)H lung MRI with lung masks:

\(^1\)H\(_1\) + Mask\(_1\)

\(^1\)H\(_2\) + Mask\(_2\)

\(\vdots\)

\(^1\)H\(_N\) + Mask\(_N\)

First, we register the lung MRI library to the lung MRI to be segmented, \(^1\)H\(_u\),

Register(\(^1\)H\(_u\), \(^1\)H\(_1\)) \(\rightarrow\) \(T_1\)

Register(\(^1\)H\(_u\), \(^1\)H\(_2\)) \(\rightarrow\) \(T_2\)

\(\vdots\)

Register(\(^1\)H\(_u\), \(^1\)H\(_N\)) \(\rightarrow\) \(T_N\)

Each calculated transformation, \(T\), registers the corresponding database image to the unsegmented \(^1\)H MRI.

The MALF algorithm takes the following input:

• \(^1\)H\(_u\)

• the transformed lung MRI, i.e.,

  • \(T_1\)(\(^1\)H\(_1\)),

  • \(T_2\)(\(^1\)H\(_2\)), \(\ldots\),

  • \(T_N\)(\(^1\)H\(_N\))

• the transformed lung MRI masks, i.e.,

  • \(T_1\)(Mask\(_1\)),

  • \(T_2\)(Mask\(_2\)), \(\ldots\),

  • \(T_N\)(Mask\(_N\))

Mask\(_u\) \(\leftarrow\) MALF(unsegmented \(^1\)H, transformed lungs, transformed masks)

Past MRI research has attempted to explore subregions of the lungs by geometric subdivisions, e.g.,

However, these subdivisions

• have no clinical consensus and

• are not based on internal anatomical structure, such as the lobes.

Given the whole lung segmentation described in the previous section, we can estimate\(^*\) a lobar division by a similar process.

which requires the use of a lobe-annotated lung library which we can derive from CT.

\(^*\) I purposely avoided the use of the word segmentation here since it might have confounding connotations.

CT lung mask library for lobe estimation

• The CT library contains only the whole lung and lobe masks.

• There is no reliable anatomical information inside \(^1\)H:

  • This makes direct segmentation difficult.

  • However, we can register the \(^1\)H and CT lung boundaries using the masks, e.g.,

\(^*\)Data and ANTs call available here.

Overview of lobe estimation in \(^1\)H MRI:

• Step 1: Segment the whole \(^1\)H lung MRI as described previously.

  • This produces an \(^1\)H lung mask.

  • The lung mask is represented as a binary image where

    • label 1 (red) = right lung

    • label 2 (green) = left lung

• Step 2: Register the CT whole lung masks to the \(^1\)H lung mask. Each CT whole lung mask is defined similarly as the \(^1\)H lung mask (i.e., label 1 = right lung and label 2 = left lung).

Overview of lobe estimation in \(^1\)H MRI (continued):

• Step 3: Transform the CT lobe masks to the \(^1\)H lung mask using the transforms derived from Step 2.

• Step 4: Use majority voting to determine the consensus labeling from the warped CT lobe masks.

• Step 5: Remove any pixels from the resulting lobe estimation which are outside the \(^1\)H lung mask.

So the results look good but

how do we quantitatively assess the proposed segmentation/estimate?

Whole lung segmentation evaluation is straightforward — just perform a leave-one out cross-validation:

• For each \(^1\)H lung MRI in the library:

  • Remove the \(^1\)H lung MRI from the library to use as the “target.”

  • Use the remaining images to segment the target image as described previously.

  • Compare the MALF segmentation with the manually annotated segmentation.

• Calculate overlap statistics over all images in the library.

Compare with previous methods I

Dice comparison with [1].

Statistics from [1] are in parentheses.

[1] Kohlmann, et al. Automatic lung segmentation method for MRI-based lung perfusion studies of patients with chronic obstructive pulmonary disease. Int J CARS 2014.

Another recent study reports a maximum Dice performance of \(~0.917 \pm 1.8\) using combined \(^1\)H and \(^3\)He images.

Compare with previous methods II

Jaccard overlap comparison with best results from LOLA11 (whole lung CT segmentation challenge)

Statistics from LOLA11 are in parentheses.

*Note that Dice \(\ne\) Jaccard which is why our numbers are not the same in both tables even though they are from the same results.

How do we evaluate \(^1\)H lobe estimation?

*Remember, I am purposely avoiding the use of the word segmentation to not confuse people as to what we are doing. We are simply making a best guess estimate of where the lobes are based on alignment of lung boundaries. Not ideal but better than the current alternative.

This is much harder to evaluate than the whole lung segmentation:

• There is little, if any, reliable internal anatomical information in \(^1\)H lung MRI which implies:

  • Manual lobe segmentation is not feasible for conventional \(^1\)H lung MRI.

  • Even image registration of MRI and CT images from the same individual are not guaranteed to be in alignment internally.

Thus, using \(^1\)H lung MRI directly for evaluation is extremely difficult (if not impossible) with our current acquisition and processing capabilities.

However,

notice something about our actual methodology, described previously:

The actual lobe estimation steps (Steps 2, 3, 4, 5) are completely agnostic with respect to the original \(^1\)H lung MRI—only the lung mask is used.

So we can revise our original evaluation question from

How do we evaluate \(^1\)H lobe estimation?

to an equivalent

How do we evaluate lung mask\(^*\) lobe estimation?

\(^*\)This is the key point: If we assume that a hypothetical \(^1\)H lung mask is equivalent to its CT counterpart, then we can perform a methodological evaluation using the CT library.

We perform a leave-one out cross-validation with CT library:

• For each Lung Mask\(_i\) (= hypothetical \(^1\)H lung mask\(_i\)):

  • Remove Lung Mask\(_i\) from the set to use as the “target.”

  • Downsample Lung Mask\(_i\) to resolution of the proton images (for improved comparability–although this step did not make much difference).

  • Use the remaining images to segment the target image as described previously (only Steps 2, 3, 4, 5).

  • Compare the Majority Voting estimation with the manually annotated lobe segmentation.

• Calculate overlap statistics over all images in the library.

Compare with related methods

Jaccard overlap comparison with best results from LOLA11 (whole lung CT segmentation challenge)

Statistics from LOLA11 are in parentheses.